Documentation

Covenant is an open, agent-native operating layer. It runs on your own machine, talks to local and remote AI agents, and provides the OS-level primitives — intent, runtime, memory, identity, permissions, comms, compositor, settlement — that humans and agents need to safely share a computer, delegate work, and pay for usage.

These docs cover concepts, architecture, reference, protocols, and operations. New here? Start with Getting started, then read Concepts for the mental model.

Browse by area

Getting started

Install the daemon, run your first agent, and submit an intent.

Concepts

Intents, agents, capabilities, memory, audit, settlement.

Architecture

How the daemon, the runtime, and the on-chain settlement program fit together.

Command-line interface

Every covenant subcommand, with arguments and exit codes.

HTTP API

Routes, request bodies, and responses on the local HTTP gateway.

Agent manifest

agent.toml schema, runtime contract, and validation rules.

Capability tokens

ed25519-signed permissions: shape, canonical encoding, verification, revocation.

MCP integration

Tool trait, native tools, and external MCP servers over JSON-RPC.

Security model

Trust boundaries, threat model, defaults, and operator responsibilities.

Status

Covenant is pre-alpha. The protocol surfaces and the local daemon are under active development; the on-chain settlement program is evolving in lock-step. We do not recommend production use yet. Feedback, sandbox experimentation, and contributions are welcome.

Where this fits in the broader stack

Covenant sits above the operating system and below user-facing agentic applications. It is not an LLM, not an agent framework, and not a chat product — it is the coordination layer those things plug into so that permissions, memory, identity, and settlement are not each application's problem to solve from scratch.